Harden eval_expr - #1744

The magnitude limit only inspects a value once it exists, so a power whose
operands are individually within the limit was computed in full before being
rejected. Evaluating 999999**999999 that way costs several seconds of CPU and
builds a number with millions of digits, which is the kind of expression the
limit was added to guard against in joblib#1744.

A power is now rejected up front when it cannot possibly fit: for any base of
magnitude 2 or more, base ** exponent is at least 2 ** exponent, so an
exponent above the bit length of the limit can never produce a value under it.
Only integer powers are checked, since a float power overflows cheaply and
already reports an error.

This also fixes the message for a large result. CPython caps integer to string
conversion at sys.get_int_max_str_digits() digits, so formatting the value
into "Numeric literal ... is too large" raised instead, and pre_dispatch of
"10**5000" surfaced a complaint about integer string conversion along with
advice to call sys.set_int_max_str_digits().

Both cases reproduce through Parallel(pre_dispatch=...).